AFAIK AMD do have a machine learning projects running.
- Caffe OpenCL GitHub - BVLC/caffe at opencl
The following work seems to not be official. But the maintainer seems to be AMD related.
- DeepCL GitHub - hughperkins/DeepCL: OpenCL library to train deep convolutional neural networks
- cltorch GitHub - hughperkins/cltorch: An OpenCL backend for torch.
hughperkins is also working on tensorflow with OpenCL.
HIP might help since it can convert CUDA code into HIP code. But , It's a pain to setup and only supports Linux. I'm still trying to setup ROCm on my GTX 970/Manjaro Linux box so I could compare the performance of HIP vs CUDA (I'm working on a GPU ray tracer).
BTW, definitely agree that AMD have to do more work on ML.